Milk mid-infrared spectra enable prediction of lactation-stage-dependent methane emissions of dairy cattle within routine population-scale milk recording schemes

Additional keywords: breeding, management, methane prediction.

Methane naturally emitted by dairy cows during anaerobic fermentation of feed is a significant greenhouse gas (Gerber et al. 2013). In addition the loss of energy as methane reduces the availability of consumed energy for the cow (Johnson and Johnson 1995). Hence, mitigating methane emissions could improve sustainability and profitability of dairy production. Factors governing the natural emission of methane from dairy cows include: diet composition (e.g. fibre, starch and lipid content), management of the herd (e.g. grassland access; Beauchemin et al. 2008) and genetics (some animals are more efficient than others and methane emissions seem to be heritable; e.g. Pickering et al. 2015). Recently, highly efficacious inhibitors of methanogenesis have also been developed (e.g. Hristov et al. 2015), which can be expected to further increase variability in methane emissions between cows and herds. Current methods to determine in vivo methane emissions (e.g. chamber, tracer gas and spot sampling techniques) are expensive and difficult to carry out on a large scale. For this reason, these direct measurement methods are unsuitable to examine the aetiology of methane emissions at a population scale. For that purpose the development of an easily measured proxy of methane emissions is desirable.

An equation has been developed to predict methane emissions from the mid-infrared (MIR) spectra of milk while accounting for lactation stage of cows (Vanlierde et al. 2015). This prediction equation could be very useful as a high throughput phenotypic screening tool within routine milk-recording programs to easily generate methane data and to undertake studies of methane emissions within- and between herds. However, accuracy and robustness of this equation can still be improved by the inclusion of new data in the calibration set to cover more variability (Vanlierde et al. 2015). This study was conducted to improve methane prediction from milk MIR spectra and to demonstrate the usefulness of the improved equation by applying it to the Walloon spectral database.

The initial equation developed in order to predict methane emissions was solely from milk MIR spectra (Dehareng et al. 2012). However, when applied to spectral data from the Walloon milk recording database, the resulting predictions of methane emissions throughout lactation (i.e. decreasing in early lactation and increasing thereafter) did not fit with the pattern expected from the literature (i.e. increasing in early lactation and decreasing thereafter; Garnsworthy et al. 2012). The initial equation was improved by replacing fixed coefficients by lactation-stage-dependent coefficients based on the use of modified Legendre polynomials (Gengler et al. 1999; Vanlierde et al. 2015). This development was based on the hypothesis that the relationship between methane emissions arising from fermentation of feed and milk MIR spectra, (which indirectly represents composition of milk) established through the equation of Dehareng et al. (2012) is also influenced by the products of fermentation processes (i.e. fatty acids; FA). Milk FA composition is dependent on the stage of lactation as rates of body tissue mobilisation differ. This strongly influences the synthesis of the different milk FA and their relative proportion in milk (Bastin et al. 2011). Changes in milk FA composition can be expected to influence the relationship between methane emissions and milk MIR spectra. For this reason, including stage of lactation in the equation enables consideration of the expected metabolic status of cows throughout lactation and was expected to improve the accuracy and robustness of the prediction (Vanlierde et al. 2015).

A total of 532 daily methane emissions measurements with corresponding daily milk MIR spectra were obtained from 165 cows using the SF₆ technique and the Fourier transform MIR spectrophotometer. Daily methane emissions within this reference dataset ranged from 180 to 802 g/day with an average ± s.d. of 430 ± 129 g/day. The reference data were highly variable, comprising data from: (1) different diets: classic barn diet with or without linseed supplement, with maize silage or fresh grass making up the greatest portion of the diet; total mixed ration; maize silage morning ration with a fibrous (straw, grass silage) evening ration; or full grazing; (2) cows of different breeds from different countries: 268 observations of Holstein cows in Belgium and 264 observations of Holstein, Jersey and Holstein × Jersey crossbred cows in Ireland; (3) different lactation parities (64 first, 43 second, 58 third or greater) and (4) different stages of lactation (18–322 days in milk). The 532 reference samples include 86 additional samples to those used by Vanlierde et al. (2015). The 86 additional samples were collected in Belgium from four different commercial farms and contribute more variability to the calibration equation through the addition of 18 new cows, four new farms and their particular management and diet systems. Moreover, those cows have been chosen particularly to have a better coverage of the entire lactation in the calibration dataset (Table 1). Indeed, Vanlierde et al. (2015) showed that for the previous version of the equation there was a lack of data at the very beginning and the end of the lactation especially for Belgian data.

Table 1.  Distribution of parity and stage of lactation of cows used to obtain new reference data in the calibration dataset (compared with the calibration dataset of Vanlierde et al. 2015) aimed to develop methane emissions prediction equation

A first derivative was applied to the milk MIR spectra. The calibration was developed from three spectral regions: between wavelengths 968 and 1577 cm⁻¹; 1720 and 1809 cm⁻¹; and 2561 and 2966 cm⁻¹ (n = 289 data points). Constant (P0), linear (P1) and quadratic (P2) modified Legendre polynomials were computed from cows days in milk at the day of methane measurement (Gengler et al. 1999) to take account of the expected metabolic status of cows. In this case Eqn 1:

where and P0, P1, and P2 were the Legendre polynomials of first, second and third order, respectively.

The calibration model was developed using first derivative MIR data points at considered wavelength × P0, first derivative MIR data points × P1, first derivative MIR data points × P2. The final modified spectra were based on 3 × 289 data points (867 data points). A modified partial least square regression strategy as implemented in the WINISI software (Foss, Hillerød, Denmark) was used.

To test the robustness of the calibration model, a five-groups cross-validation was done: reference data were divided randomly into five groups, five calibration models were developed by removing each time one group that is predicted by the four others.

Predictions of methane emissions were obtained by applying the prediction equation to milk MIR spectra collected between January 2007 and January 2015 from 1675 herds in the Walloon Region of Belgium. Only spectra collected between 5 and 365 days in milk were kept. To discard spectra too spectrally different from spectra used for the calibration process, standardised Mahalanobis distance (global H distance; GH) between each spectra and spectra used for the calibration was calculated and spectra with a GH value greater than 3 were deleted from this study (e.g. Shenk and Westerhaus 1991; Soyeurt et al. 2012). Methane prediction lower than 150 g/day and higher than 950 g/day were considered as outliers (Vanlierde et al. 2015) and were excluded from this study (0.07% of the 2 049 366 spectra with a GH value lower than 3). A total of 2 047 893 MIR spectra were used from 145 249; 122 861; and 86 725 Holstein cows in first, second or third lactation, respectively. The methane prediction equation was applied to these spectra along with the lactation stage of the cows. Descriptive statistics were developed based on these predictions in order to demonstrate the usefulness of the methane prediction equation.

The predicted methane emissions (g/day) obtained with the equation based on 532 reference data [versus those obtained with the one based on 446 reference data (Vanlierde et al. 2015)] had a calibration coefficient of determination (R²c) of 0.74 (vs 0.75), a cross-validation coefficient of determination (R²cv) of 0.70 (vs 0.67), a standard error of calibration (SEC) of 66 g/day (vs 63) and a standard error of cross-validation (SECV) of 70 g/day (vs 73). The predicted methane values ranged from 168 to 715 g/day with an average ± s.d. of 430 ± 111 g/day (vs 415 ± 127 g/day), which was similar to the reference values. The statistical results between the two versions of the equation are similar; however, the new equation is more robust as evidenced by the smaller difference between R²c and R²cv of the new equation as well as a smaller difference between SEC and SECV.

The relationship between predicted and measured methane emissions is shown in Fig. 1. It is clear from the linear relationship that MIR prediction was highly correlated with actual methane emissions and that it is possible to identify cows with high or low daily emissions. However, the accuracy and robustness of this prediction equation could be further enhanced by the addition of new reference values (i.e. methane measurements and milk MIR spectra with variability not already included in the model). Such new reference values should be added to the calibration dataset to increase coverage of population variability.

Fig. 1.  Relationship between the predicted methane emissions determined from mid-infrared spectra using methane prediction equation and the measured methane emissions of cows from the calibration dataset used to develop methane prediction equation (n = 532).

The developed prediction equation for methane emissions was applied to milk MIR spectra to demonstrate its potential usefulness. The proportion of spectra with GH values higher than 3 excluded from the spectral database was lower compared with that obtained by Vanlierde et al. (2015; 5% vs 14%, respectively), indicating that robustness of the prediction equation was improved in this study by covering more spectral and methane variabilities. Averages ± s.d. of methane predictions were 443 ± 69 g/day, 468 ± 74 g/day, and 475 ± 75 g/day for cows in first, second, and third lactation, respectively. As expected, multiparous cows had greater emissions than primiparous cows. The quantity of feed eaten has a large influence on methane emissions (e.g. Pickering et al. 2015) and on average the feed intake of multiparous cows is greater than primiparous cows. Predictions of average daily methane emissions throughout lactation are shown in Fig. 2.

Fig. 2.  Predicted daily methane emissions from milk mid-infrared (MIR) spectra throughout lactation averaged for first, second and third parity cows in the Walloon milk recording database (n = 857 710, n = 700 597, and n = 489 586 cows in first, second, and third lactation, respectively).

Predicted methane emissions showed increasing values from the beginning of lactation until ~100 days in milk and then decreasing values from 200 days in milk until the end of lactation as expected (e.g. Garnsworthy et al. 2012). As found for averages of methane predictions, the change in predicted methane emissions during lactation was less for cows in first lactation compared with cows in second and third lactation. Based on methane predictions, seasonal variation (e.g. monthly and yearly) of average methane emissions were also studied. Predicted methane emissions were on average lower during the summer season compared with other seasons although there were less data in July (Fig. 3). This decrease in predicted methane emissions could be explained by a decrease in feed intake when temperature increases during summer (Hammami et al. 2015) and also because the majority of dairy cows in the Walloon Region of Belgium are outdoors grazing in the summer period. In general, dry matter intake of grass-fed cows is expected to be lower compared with cows fed with a total mixed ration (Robertson and Waghorn 2002) leading to a decrease of methane emissions (O’Neill et al. 2011) because dry matter intake represents the main driver of methane production (Hegarty et al. 2007). The majority of cows used in this study calved during autumn (36% of cows calved in September, October, or November). Therefore, the observed decrease in predicted daily methane emissions during the summer period could also be due to a decrease in feed intake, which commonly occurs at the end of lactation (Zom et al. 2012). Predicted daily methane emissions also differed between years (Fig. 4). On average, cows produced less methane from 2011 to 2014 than earlier years. Many factors could explain these differences such as variation of temperature, quality of forage, and different animals. Further investigations are needed to have a better understanding of these differences. The same trends observed in Fig. 3 and Fig. 4 (i.e. lower methane production during summer and from 2011 to 2014) are shown in Fig. 5 although amplitudes of variation differ by month and year.

Fig. 3.  Seasonal variation in average daily methane emissions of dairy cows (with relative standard deviation) predicted from milk mid-infrared (MIR) spectra by month. Points represent monthly averages drawn from the following milk MIR observations (n = 186 215 for January; n = 166 409 for February; n = 181 938 for March; n = 185 120 for April; n = 194 317 for May; n = 170 012 for June; n = 62 655 for July; n = 158 124 for August; n = 170 850 for September; n = 192 156 for October; n = 195 576 for November; and n = 184 521 for December).

Fig. 4.  Annual variation in average daily methane emissions of dairy cows (with relative standard deviation) predicted from milk mid-infrared (MIR) spectra by year from 2007 to 2014. Points represent annual averages drawn from the following milk MIR observations (n = 147 383 for 2007; n = 222 820 for 2008; n = 236 753 for 2009; n = 316 147 for 2010; n = 303 180 for 2011; n = 298 420 for 2012; n = 275 007 for 2013; n = 241 015 for 2014).

Fig. 5.  Average daily methane emissions of dairy cows predicted from milk mid-infrared (MIR) spectra in the Walloon milk recording scheme between 2007 and 2015 (n = 2 047 893 milk MIR spectra).

Regional differences in methane emissions of dairy cows were also identified. The two most extreme differences, considering a period of 7 years (from August 2007 to December 2014), were found between the average daily methane emissions of cows located in the town of Doische within the region of the Province of Namur was 422 ± 74 g/day to the cows located in the town of Attert within the region of the Province of Luxembourg emitting 482 ± 68 g/day on average. Without more detailed information from the field the increased daily methane emissions of the cows from this town in the Province of Luxembourg region cannot yet be precisely explained. However, the standard deviations of average daily methane emissions were in the same range as the SECV of the prediction equation.

Thanks to this tool and some of these descriptive statistics, management strategies aimed at reducing methane emissions or increasing the feed conversion efficiency of dairy cows could be developed. Indeed, MIR spectra could be routinely recorded within milk-recording schemes (Friedrichs et al. 2015) allowing large-scale prediction of methane emissions. Screening of cows’ methane emissions between herds could be performed. Predictions of methane emitted by cows could be also evaluated within a herd. For example, the daily methane emissions predicted from milk MIR spectra throughout lactation for two cows within the same herd is shown in Fig. 6. Differences between methane predictions of these cows could be explained by different factors. One of these factors should be the calving season. Indeed, cow A calved during autumn whereas cow B calved during spring. Therefore, peak lactation for cow B occurred at the beginning of summer following, which the feed intake of cow B was expected to decrease and this feed intake decrease should lead to a decrease in methane emissions. In comparison, peak lactation for cow A occurred during winter with no decrease of feed intake expected. A comparison between the average predicted methane emissions of first lactation cows in 30 Walloon herds with relative standard deviation is shown in Fig. 7. Huge differences of average predicted methane emissions between these herds could be observed. These differences between herds could be related to herd characteristics such as diet composition and management. Further investigations are needed to explain these differences.

Fig. 6.  Predicted daily methane emissions from milk mid-infrared (MIR) spectra throughout lactation for a cow calving in either autumn (cow A) or spring (cow B). Both primiparous cows were in the same herd.

Fig. 7.  Averaged predicted methane emissions from milk mid-infrared (MIR) spectra (with relative standard deviation) of dairy cows in first lactation from 30 Walloon herds (n = 492 for herd 1; n = 1562 for herd 2; n = 489 for herd 3; n = 770 for herd 4; n = 403 for herd 5; n = 873 for herd 6; n = 501 for herd 7; n = 850 for herd 8; n = 891 for herd 9; n = 825 for herd 10; n = 602 for herd 11; n = 516 for herd 12; n = 819 for herd 13; n = 677 for herd 14; n = 1283 for herd 15; n = 682 for herd 16; n = 646 for herd 17; n = 604 for herd 18; n = 325 for herd 19, n = 362 for herd 20; n = 310 for herd 21; n = 820 for herd 22; n = 557 for herd 23; n = 739 for herd 24; n = 2238 for herd 25; n = 887 for herd 26; n = 371 for herd 27; n = 846 for herd 28; n = 397 for herd 29; n = 456 for herd 30).

Milk MIR spectroscopy now provides a very powerful, low cost and high throughput (population scale) screening method to identify cows within national herds that display a low methane emissions phenotype. The identification of genetic markers common to these animals could enable genetic or genomic selection for reduced enteric methane emissions. Such breeding objectives are desirable as methane emissions represent a substantial economic loss of dietary energy, and this trait appears to be variable and heritable (e.g. Pickering et al. 2015). Genetic correlations of methane emissions with other traits of interest such as milk, fat and protein yield, could be estimated simultaneously as MIR spectra are recorded within milk-recording schemes.

The prediction of daily methane emissions of dairy cows using MIR spectrometry analysis that could be already available in routine within milk recording represents a powerful and low cost tool to compare methane emissions between cows, herds, years, diets and management practices. Cows with high milk production and low methane emissions could be distinguished. And there are many more applications of this tool. However, additional methane measurements are needed in the calibration dataset to improve accuracy and robustness of this tool. Moreover, through application of the technology we have developed, other investigations should be achieved in order to have a better understanding of the parameters which influence enteric methane emissions and genetic selection could be undertaken while taking into account other important production parameters such as milk yield, fertility, and efficiency. Inclusion of MIR-derived methane emissions traits in genetic selection indices is expected to result in an incremental and permanent improvement in milk production efficiency and greenhouse gas emissions over time.